The Industrialist's Case for Applied AI

Executive framing: from AI rhetoric to industrial economics

Applied AI is best understood as an industrialization curve, not a software moonshot. The important question for investors is no longer whether models can generate impressive outputs on a screen; it is whether those outputs can be embedded into recurring workflows, governed inside operating systems, and translated into measurable economics. That distinction matters because the value of AI in industry is likely to accrue less to generic model access than to the companies that own the work itself: the process steps, the decision rights, the data exhaust, and the capital allocation levers that turn insight into action.

This is why the familiar “operators versus adopters” framing is useful, but incomplete. In practice, the real divide is between businesses that can operate AI and those that merely adopt it. Operators have proprietary workflows with enough repetition to learn from, enough scale to amortize implementation costs, and enough management discipline to convert recommendations into changed behavior. Adopters may have access to the same models, but not the same decision latency, process standardization, or feedback loops. In industrial settings, those frictions can matter more than model capability itself.

The early evidence supports a measured view. McKinsey’s recent operations research suggests AI investments in manufacturing and back-office operations are beginning to pay back faster, but that the best outcomes are concentrated in leading organizations rather than spread evenly across users.^[1] The same research emphasizes that COOs need the right operating structure, data governance, and change management to turn gen AI and agentic AI into impact, not just prototypes.^[2] In other words, the economic advantage does not come from “having AI” in some abstract sense. It comes from shortening the distance between signal and decision, and between decision and execution.

That is the lens we should use for industrial investing. AI is most valuable where it reduces decision latency in high-frequency, economically material workflows: maintenance, quality, procurement, scheduling, customer service, engineering, and sales support. The prize is not just labor substitution. It is fewer breakdowns, faster throughput, lower scrap, improved service levels, and better capital allocation. The businesses that can institutionalize those gains are the ones most likely to convert AI into margin expansion and productivity compounding.

The rest of the article will avoid hype and focus on what can be observed: where ROI is showing up, what separates pilot-stage success from durable operating improvement, and why industrial heritage may matter as much as technical ambition. In this market, the winning question is not “which model is best?” but “which business can make AI part of how work gets done?”

Why industrial AI resembles a familiar value-creation curve

Industrial technology tends to create value in a familiar sequence: first as a novelty, then as a point solution, and finally as a production capability that is embedded into recurring workflows. The history of factory automation, ERP, lean process redesign, and advanced analytics all points to the same lesson: the durable gains do not come from owning the tool in isolation, but from reorganizing work around it. AI is moving through that same curve. The companies that will matter most are not necessarily those with the largest model budgets, but those that can turn model output into routine operational decisions, with clear accountability and measurable economics.^[5]

That distinction matters because AI access is becoming less scarce. Model quality is improving quickly, and many general-purpose capabilities are increasingly available through cloud platforms and software vendors. In that environment, generic model capability is a necessary input, but it is not a moat by itself. The harder problem is organizational: identifying repetitive decisions, wiring AI into the systems where those decisions are made, and then governing the handoff between human judgment and machine recommendation at scale. McKinsey’s work on operations leaders suggests the operating frontier is moving from experimentation to deployment, with leading organizations already pulling ahead on AI-enabled operations rather than merely piloting use cases.^[1]

That pattern is also visible in the Global Lighthouse Network literature. McKinsey notes that many manufacturers were stuck in “pilot purgatory” in the late 2010s, with more than 70 percent reporting pilots that failed to generate significant business impact; the winners were the firms that built the capabilities to deploy across factories and supply networks, not just prove concepts in isolated teams.^[5] In other words, the bottleneck was never simply the algorithm. It was the ability to standardize the workflow, own the process data, and scale adoption across similar decisions.

This is why AI should be analyzed less like a software product cycle and more like an industrial industrialization curve. The first returns come from isolated pockets of productivity. The durable returns come when AI becomes part of the operating system: embedded in scheduling, quality checks, maintenance planning, procurement, customer support, and engineering change management. At that stage, the economic question shifts from “Can the model do this task?” to “Can the business repeatedly capture the value, measure it, and compound it?”

The most important implication for investors is that AI value is likely to accrue where workflow ownership, process discipline, and decision latency are already strategic variables. A firm with proprietary production data and tightly managed operations can convert AI into margin expansion, lower downtime, and better capital allocation. A firm without those attributes may still use AI, but it is more likely to rent intelligence than to own the returns.

Where the measurable ROI is appearing first

The first measurable returns from applied AI are showing up where industrial businesses already have a lot of repetitive decisions, high-frequency exceptions, and expensive latency: maintenance, quality, forecasting, scheduling, customer service, procurement, and engineering support. That is not where the hype is loudest, but it is where the economics are clearest. McKinsey’s operations research in 2025 finds leading companies are already using AI to improve supplier negotiations, strengthen quality control in maintenance, and automate parts of sales and service work that were previously too labor-intensive to scale.^[1][2]

The pattern matters. In most of these use cases, AI is not “replacing judgment” so much as compressing cycle times and reducing avoidable friction in an existing decision loop. Predictive maintenance reduces unplanned downtime by identifying failure risk earlier; visual inspection reduces scrap and rework by catching defects faster; forecasting and planning reduce stockouts and excess inventory; customer service tools reduce response latency and handle more long-tail interactions; procurement tools speed supplier analysis and negotiation; and engineering copilots reduce the time spent searching, drafting, summarizing, or checking work.^{[1][2][3][4][7]}

Function	Typical AI use	Primary value pool	Evidence point from recent industry research
Maintenance	Predictive analytics, repair copilots, parts planning	Downtime reduction; labor productivity	McKinsey cites early adopters using gen AI to improve equipment maintenance quality control and to support aircraft MRO workflows under labor constraints, where keeping assets available is the core economic objective.[2][7]
Quality control	Computer vision, defect detection, root-cause support	Scrap, rework, warranty reduction	Operations leaders are applying AI in manufacturing and maintenance functions; the reported value comes from earlier detection and better decision support, not from fully autonomous inspection.[1][2]
Procurement	Spend analytics, supplier intelligence, negotiation support	Cost avoidance; supply continuity	McKinsey’s procurement research highlights better data and AI-enabled sourcing decisions across category strategy, supplier assessment, negotiation, and ongoing supplier performance management.[3][4]
Customer service / sales support	Agent assist, chatbots, account triage	Response latency; revenue capture	One digital marketing platform used gen AI to manage “long tail” sales accounts that were previously too labor-intensive, producing an annual revenue gain of more than $30 million.[2]
Engineering productivity	Drafting, search, summarization, technical copilot tools	Throughput; cycle time; labor leverage	Airline maintenance and broader operations research emphasizes productivity gains from copilots that reduce clerical and information-search work, freeing skilled labor for higher-value tasks.[7]

Procurement is a particularly good example of where AI can create real, if sometimes underappreciated, enterprise value. In volatile markets, purchasing decisions depend on demand signals, supplier behavior, lead times, and contract details that are hard to reconcile manually at speed. McKinsey’s 2024 procurement work argues that better data and AI can improve decisions across the sourcing life cycle, from category strategy to supplier performance management.^[3][4] The return is often not just lower price; it is better continuity of supply, fewer expedites, and less time spent on low-value analysis. For industrial operators, that can matter more than the model sophistication itself.

Customer service and sales support have also produced visible early wins because the work is text-heavy, repetitive, and measurable. When AI can resolve standard queries, route exceptions, or help manage smaller accounts economically, the P&L impact shows up as lower cost-to-serve and incremental revenue from segments that were previously uneconomic to cover manually.^[2] The strongest results usually come from augmenting front-line teams rather than bypassing them entirely.

What is notable is where the gains are not coming from first. In industrial settings, the most reliable ROI is rarely from a dramatic “zero human” transformation. It is from taking minutes, hours, and days out of recurring processes, and from reducing the error rate in decisions that already occur thousands of times a month. That is why the early winners tend to look less like software demos and more like better operations: fewer outages, fewer defects, faster quote turnaround, tighter inventory, and less idle labor.^[1][2][7]

Pilot gains versus scaled operating improvements

The most common mistake in industrial AI is to mistake a good pilot for a good business. A model can impress in a controlled workflow, yet leave the P&L unchanged if the output is not wired into planning, maintenance schedules, procurement approvals, pricing, or frontline incentives. McKinsey’s operations research notes that many manufacturers spent years in “pilot purgatory,” with use-case trials that did not create meaningful business impact until companies built the capabilities to deploy at speed and scale.^[5]

That distinction matters because the economics are different. Pilot gains are often real but local: a chatbot reduces average handle time in a contact center; a vision model flags defects on one line; a maintenance model improves a single asset class. Scaled operating improvements show up elsewhere: lower downtime across a plant network, fewer stockouts from better forecasting and scheduling, lower scrap and rework from closed-loop quality control, and faster decision-making because exceptions are triaged automatically rather than manually. In McKinsey’s newer Lighthouse work, the point is explicit: when technologies mature, the challenge becomes speed and scale, and the factory or supply network—not the isolated use case—becomes the pilot.^[5]

Deployment stage	Typical evidence of success	What it usually does not prove	Investor question
Pilot / proof of concept	One workflow, one site, one team; KPI improvement over weeks or months	Replicability across plants, regions, or business units	Can it survive integration, governance, and frontline use?
Scaled deployment	Multiple sites, standardized process, monitored performance drift	Enterprise-wide margin impact if incentives and systems remain unchanged	Is the model embedded in the operating cadence and decision loop?
Operating improvement	Sustained reduction in downtime, scrap, cycle time, or service latency across the base	That the benefit was “just AI” rather than process redesign plus adoption discipline	Does the improvement persist without constant vendor intervention?

The gap between pilot and scale is usually not about model quality alone. It is about the unglamorous work of integration: connecting AI to legacy ERP and MES systems, defining accountable human overrides, monitoring drift, training operators, and deciding who owns the savings. McKinsey’s Lighthouse work shows that leading manufacturers are increasingly doing this by coupling AI with broader 4IR capabilities; nearly 60% of the top use cases implemented by the newest Lighthouses relied on AI technologies, but the value came from deployment capability, not from the model in isolation.^[8]

For investors, the practical implication is straightforward. Do not underwrite AI as a one-off productivity tool. Underwrite it as a change to the operating system. If the company can convert a single successful workflow into a repeatable management process, the upside can compound. If it cannot, the demo may still be attractive—but the returns will likely stay trapped inside the pilot.

The operating prerequisites for durable AI value capture

The question for industrial investors is not whether AI is impressive; it is whether the business has the operating conditions to turn AI into repeatable economic gain. The companies most likely to do so share a few traits: processes that recur often enough to learn from; standardized workflows that can be instrumented; clean, accessible data; and a management system that can act on AI outputs quickly, not quarterly. McKinsey’s work with operators and manufacturers points in the same direction: the highest-impact deployments are not one-off demos, but capabilities embedded into production networks, procurement, quality, and maintenance routines, with speed and scale as the real test.^[2][5][8]

That matters because AI compounds only when it sits inside a decision loop. A model can flag a maintenance issue, suggest a purchase timing, or draft a customer response—but the value is realized only if someone, or something, changes a work order, a sourcing decision, a dispatch sequence, or a pricing action. Businesses with many similar decisions can amortize the setup cost across dozens of plants, product lines, or customer segments. Businesses with fragmented sites, bespoke operating rules, or poor master data often discover that each deployment is effectively a new implementation.

In practice, the operating prerequisites are less glamorous than the headlines suggest:

• Repeatability: the workflow must recur frequently enough for the system to learn and for management to measure drift.
• Standardization: similar tasks should be performed in similar ways across sites or teams; otherwise AI scales only as a consulting project.
• Data accessibility: the relevant signals—spend, machine telemetry, quality outcomes, service tickets, or engineering documents—must be available in usable form.
• Decision ownership: there must be a clear operator accountable for acting on the recommendation and for capturing the savings or throughput gains.
• Economic pain: the use case should target a real bottleneck, such as downtime, scrap, expedite cost, stockouts, or long sales-cycle friction.

Procurement is a good example of why this matters. McKinsey notes that procurement sits at the confluence of large internal and external data sets and that better data can improve sourcing strategy, supplier assessment, negotiation, and ongoing supplier performance management.^[4] But that only translates into durable value if the organization actually uses those insights to change buying behavior, contract terms, and approval discipline. Otherwise, AI becomes another layer of analytics sitting above the same old process.

Manufacturing “Lighthouses” illustrate the difference between isolated use cases and enterprise capability. McKinsey reports that earlier waves were stuck in “pilot purgatory,” while later leaders built the capabilities to deploy AI and other 4IR tools with speed and scale; it also notes that nearly 60 percent of the top use cases implemented by the 21 newest Lighthouses relied on AI technologies.^[5][8] The implication for investors is straightforward: the best AI-ready businesses are not simply those with the most ambitious pilots. They are the ones with the operating discipline to turn a few wins into a system-wide playbook.

AI readiness factor	What investors should look for	Why it matters for durable ROI
Process repeatability	High-frequency workflows; similar decisions across sites, shifts, or customers	Enables learning, benchmarking, and reuse of models and controls
Data quality and access	Connected systems, clean master data, accessible historical outcomes	Reduces time spent reconciling data and improves model usefulness
Workflow standardization	Common SOPs, shared KPIs, consistent exception handling	Makes deployment scalable beyond a single pilot site
Management cadence	Regular review of AI recommendations and outcome tracking	Converts recommendations into action and captures savings
Scale of pain point	Material downtime, scrap, expedite spend, service backlog, or cycle-time drag	Supports a large enough benefit to justify integration and change costs

The investor’s practical takeaway is that AI readiness is not a slogan; it is an operating profile. Firms that combine recurring decisions, accessible data, standardized execution, and strong management cadence are positioned to compound small gains into margin expansion, faster throughput, and better capital allocation. Firms that lack those traits may still use AI, but they are less likely to own the economics of it.^[2][5][8]

Where value accrues in the industrial AI stack

The economics of applied AI are starting to look less like a software adoption story and more like an industrial stack story. In manufacturing and other asset-intensive businesses, the value does not automatically accrue to whoever buys the model or builds the flashiest demo. Some rents will be captured upstream by cloud, semiconductor, and model infrastructure providers; some by industrial software vendors and systems integrators; and only a portion will remain with the operator that actually owns the workflow and the P&L. That split matters for investors because the highest-return layer is not always the layer that headlines the most.

The recent evidence is consistent with that view. McKinsey’s 2025 operations research finds that leading organizations are already generating payback from AI in operations faster than the rest, while also widening the gap between “accelerators” and everyone else.^[1] McKinsey’s Lighthouse work makes the same point more forcefully: when technologies reach maturity, the bottleneck becomes speed and scale, not whether the use case exists at all, and the factories and supply networks themselves become the pilots.^[5] In other words, value creation shifts from model novelty to operating system redesign.

That is where the stack becomes important. Infrastructure providers can monetize compute intensity; software incumbents can monetize workflow embedding; and systems integrators can monetize implementation complexity. These are real businesses, and in some cases they may capture more durable economics than the end user, especially where deployment requires specialized integration, governance, and change management. But the operator still has a distinctive advantage when it controls the data exhaust, the decision loop, and the cadence of execution. In those settings, AI becomes less a standalone product and more a compounding capability inside scheduling, maintenance, quality, procurement, or service processes.^[8]

That also means the most defensible value is often not in the generic model itself, which is likely to commoditize over time, but in the proprietary workflow layer: the permissions, interfaces, feedback loops, and historical process data that determine whether the system actually changes behavior. A company with standardized operations and clean process data can route AI into recurring decisions; a fragmented operator with inconsistent systems may simply create another dashboard. The economic rent follows the ability to turn intelligence into action, not the ability to generate an answer.

For investors, the practical implication is straightforward: underwrite the layer that owns the workflow, not just the layer that talks about AI. In some cases that will be the end operator; in others, the better economics will sit with the software vendor, cloud platform, or integrator enabling the deployment. The question is not whether AI is valuable. It is where, in the stack, that value is retained after implementation friction, model substitution, and operating complexity are paid for.

Counterarguments and failure modes

The bullish case for industrial AI is real, but it is not linear, and it is not automatic. The same forces that make AI attractive—rapid deployment, low marginal inference cost, and broad applicability across functions—also make it easy to overstate near-term economic impact. In practice, value can be diluted by model commoditization, implementation friction, weak data foundations, and the simple fact that many industrial organizations do not have the operating discipline to turn a better algorithm into a better P&L. McKinsey’s recent work on operations is explicit that the companies capturing value are those that define the right operating structure, data governance model, and change-management approach; without those capabilities, AI remains a tool, not an operating advantage.^[2]

The first risk is that model capability itself gets competed away faster than investors expect. Foundation models, copilots, and agents are improving quickly, but for many industrial use cases the model is not the scarce asset. The scarce asset is the workflow: the decision sequence, exception-handling logic, approval rights, and feedback loop that determine whether the output changes anything operationally. If a competitor can access similar models through cloud providers, enterprise software, or a system integrator, then the economic moat shifts away from “having AI” toward “having proprietary processes and process data.” That cuts both ways. It means end operators can win, but it also means some value will accrue to infrastructure and software incumbents that sit closer to the tools and distribution rails than the plant floor or service desk.^[5][8]

A second failure mode is implementation risk. McKinsey notes that many manufacturers historically got stuck in “pilot purgatory,” where use cases worked in a controlled setting but failed to scale into broad business impact.^[5] That gap is often caused by integration work that is less glamorous than the model demo: connecting ERP, MES, CMMS, quality systems, and frontline workflows; creating a clean training set; building human override logic; and maintaining performance over time. If an AI system only helps a subset of users, or only works when a few experts babysit it, the economics can look good in a presentation and weak in the operating statement.

Common AI failure mode	Typical operational effect	Why returns disappear	Investor diligence focus
Model commoditization	Competing firms get similar performance	No durable pricing power	Is the moat in workflow/data, not the model?
Integration and change failure	Pilot works, enterprise adoption stalls	Value stays local	Can the solution be embedded into core systems and incentives?
Cyber and data exposure	New attack surfaces and leakage risk	Costs rise faster than savings	How is access controlled, audited, and segmented?
Regulatory / safety constraints	Human review remains mandatory	No full automation of the decision loop	Where is AI advisory versus decision-making?

Cybersecurity is another material downside. Industrial AI typically increases the number of connected systems, data pathways, and third-party dependencies. That can improve visibility, but it can also widen the attack surface. In regulated or safety-critical settings, the cost of a mistake is asymmetric: a small productivity gain is not worth a material increase in operational or compliance risk. Similarly, labor resistance can blunt adoption when frontline teams see AI as surveillance, deskilling, or a headcount reduction program rather than a productivity aid. Even when the technology is sound, the change-management burden can be large enough to delay or diminish returns.^[2][5]

There is also a governance problem: in industrial environments, accountability cannot be outsourced to the model. If AI recommends a maintenance intervention, a procurement action, or a production change, management still needs explicit rules for who approves, who overrides, and how performance is measured. Where accountability is unclear, organizations often default to caution, which means the AI becomes an advisory layer rather than a lever for throughput, downtime reduction, or capital allocation improvement.

The practical conclusion is not to avoid AI, but to underwrite it conservatively. The strongest cases are those where value is measurable, the process is repeatable, the data is owned internally, and management can name the exact operating decision the system will improve. If those conditions are absent, AI is less likely to be a margin expansion engine than an expensive experiment that leaks value to vendors, integrators, and internal complexity.

A practical diligence framework for investors

An effective AI diligence process starts by asking a narrower question than “Is management using AI?” The better question is whether AI can be embedded into the company’s recurring operating decisions often enough, with enough data quality, to change unit economics. In industrial businesses, that means looking for workflows with high decision frequency, measurable pain, and a clear path from recommendation to action. McKinsey’s recent operations work points to the same conclusion: leading companies are no longer treating AI as isolated pilots, but as an operating capability that must be governed, integrated, and scaled across factories, supply networks, and back-office processes.^[1][5]

A practical scorecard can be built around eight questions. First, is the underlying process repeatable, or is it so bespoke that every intervention becomes a custom consulting project? Second, is the data available, timely, and structured enough to support reliable inference? Third, are the decisions frequent enough that small improvements compound? Fourth, is there clear economic pain—downtime, scrap, rework, expediting, missed service levels, slow turns, or labor bottlenecks—so value can be measured rather than inferred? Fifth, does management own the use case, or is AI delegated to an innovation team with no P&L accountability? Sixth, is there a defined measurement framework, with baseline performance, control groups where possible, and a plan to track post-deployment drift? Seventh, can the business integrate AI with legacy systems and frontline workflows without creating parallel processes? Eighth, does the company have acceptable cyber and governance controls for more connected, decision-support systems?^[2][5]

AI readiness factor	What to verify	Why it matters	Red flags
Data quality	Clean, accessible process data; labeled outcomes; consistent definitions	Determines whether models can improve decisions at scale	Spreadsheet archaeology; disconnected systems; manual reconciliation
Workflow standardization	Similar tasks repeated across sites, lines, customers, or suppliers	Allows one deployment to compound across many decisions	Every plant or team runs a different process
Decision frequency	Daily, hourly, or event-driven decisions with measurable consequences	Creates enough repetitions for AI gains to become material	Rare decisions with limited feedback loops
Implementation discipline	Owner, cadence, training, exception handling, and accountability	Turns a pilot into a production system	Innovation theater without line management ownership
Legacy integration	Ability to connect to ERP, MES, CMMS, CRM, or procurement tools	Prevents AI from sitting outside the operating loop	Standalone dashboards that never affect action

Investors should also distinguish where value is most likely to show up. In maintenance, quality control, and procurement, the strongest cases usually come from fewer disruptions, less scrap, and better buying decisions, which flow through the income statement relatively quickly.^[2][4] In customer service, sales support, and engineering productivity, the first benefit may be higher throughput per employee or faster response times, with EBIT improvement depending on whether management actually resets staffing or redeploys capacity.^[1][2] The key diligence point is simple: if AI is only improving “visibility,” but not changing cadence, staffing, inventory, or capex choices, then the value is likely overstated.

For underwriting, the best businesses are not those with the loudest AI narrative. They are the ones where AI can be measured in fewer outages, shorter cycle times, lower working capital, and better capital allocation. That requires operator-level evidence, not demo-level enthusiasm. A company that can show recurring adoption, process control, and verified ROI is more investable than one that can only show a pilot deck.

For Welkin, operating heritage matters because it sharpens that distinction. It helps identify whether a deployment is actually changing the system of work, whether savings are durable, and whether the organization can absorb the change. In other words, the edge is not in believing more in AI; it is in asking better questions about how AI becomes operating leverage.

Where value creation is most likely by function

The industrial AI opportunity is easiest to understand through functions, not slogans. In operations-heavy businesses, the near-term value is not in “deploying AI” in the abstract; it is in removing friction from a handful of recurring decision loops that already have economic owners, measurable service levels, and existing budgets. McKinsey’s recent COO and procurement work points to the same pattern: early adopters are already using gen AI to improve supplier negotiations, quality control, and maintenance, while some organizations are using it to serve low-touch sales accounts that were previously uneconomic to cover.^[2] The practical implication is straightforward: the first wave of returns should show up where AI shortens cycle times, reduces exceptions, and improves decision consistency in high-volume workflows.

Maintenance is often the cleanest route to EBIT improvement because the link between better prediction and lower cost is direct. Predictive analytics and gen AI can help prioritize work orders, diagnose failure modes, and assist technicians with repair instructions, with the economic payoff coming from fewer unplanned outages, lower overtime, and higher asset availability.^[2][7] In asset-intensive businesses, that can mean more throughput from existing capex rather than a new spending program. The same logic applies to quality control: visual inspection and anomaly detection can reduce scrap, rework, and warranty leakage, but only if the model is embedded into the inspection and escalation process rather than treated as a sidecar dashboard.^[2]

Procurement is another early winner, but the value often lands in working capital and margin protection before it appears in headline growth. Better spend analytics, supplier intelligence, and negotiation support can sharpen sourcing decisions, particularly where markets are volatile and inputs are differentiated.^[3][4][6] The real prize is not a one-time cost takeout; it is a procurement function that can respond faster to price swings, shortages, and specification changes, preserving service while avoiding excess inventory or emergency buys.

Function	Typical AI use case	Primary economic impact	Value tends to show up first in
Maintenance	Predictive maintenance, repair copilots, work-order triage	Higher uptime, lower downtime, lower overtime	EBIT and throughput
Quality control	Visual inspection, defect detection, root-cause support	Less scrap, less rework, fewer warranty claims	EBIT and service quality
Procurement	Spend analytics, supplier screening, negotiation support	Lower input cost, lower inventory, better supply continuity	Margin and working capital
Customer service	Agent assist, self-service resolution, case summarization	Lower cost-to-serve, faster response times	EBIT and service levels
Sales enablement	Lead prioritization, account coverage, proposal drafting	More coverage of low-touch accounts, higher rep productivity	Revenue and sales efficiency
Engineering productivity	Drafting, code/technical document support, design search	Faster iteration, lower non-value-added time	Throughput and project cycle time
Scheduling / planning	Demand forecasting, production sequencing, labor planning	Higher utilization, fewer stockouts, better on-time delivery	Throughput and working capital

Customer service and sales enablement tend to generate visible productivity gains quickly because the work is text-heavy, repetitive, and measurable. AI can absorb routine inquiries, summarize cases, draft responses, and help sales teams cover accounts that were previously too small or too complex to serve efficiently.^[2] But the best results usually come when management changes coverage models, routing rules, and incentive plans; otherwise, the benefit stays trapped as “time saved” instead of operating profit.

Engineering productivity and scheduling can also create meaningful value, though often less visibly. In engineering, the gain is usually faster design cycles and less manual documentation. In planning and scheduling, the benefit is higher asset and labor utilization, fewer bottlenecks, and better delivery reliability. These are classic industrial levers: AI is useful when it improves the cadence of decisions, not when it merely adds another report.

The common thread is that the highest-conviction use cases are not those where AI replaces judgment wholesale. They are the ones where it reduces exceptions, compresses latency, and improves the quality of recurring decisions across enough volume to matter.

What Welkin’s operating heritage changes

Welkin’s operating heritage matters because AI value in industry is rarely created by picking the best model in the abstract; it is created by knowing which workflows are repetitive, which constraints are real, and which “transformation” claims will never survive contact with a plant, a procurement desk, or a field service organization. The recent evidence base points in the same direction: leading operators are no longer treating AI as a standalone experiment, but as a production capability that must be governed, scaled, and embedded into operations with speed and discipline.^[1][5]

That is where operating experience becomes an underwriting edge. A team with hands-on exposure to throughput, downtime, yield, changeovers, and working-capital constraints is better equipped to tell the difference between a useful pilot and a durable process improvement. It is easier to see whether a maintenance model actually reduces unplanned stops, whether a quality tool lowers scrap without adding a new review bottleneck, or whether a sales assistant simply shifts activity without changing conversion economics. McKinsey’s recent work on operations leaders and manufacturing Lighthouses underscores that the companies pulling ahead are the ones that can move from isolated use cases to speed and scale—meaning they build the operating structure, data governance, and change-management muscle to make AI part of the system, not a side project.^[2][5][8]

For diligence, that translates into a more skeptical and more practical posture. Operating heritage helps us ask: What is the decision being improved? How often does it recur? Where does the data come from, who owns it, and what happens when the model is wrong? Is the savings measurable in P&L terms, or merely anecdotal? Can the workflow absorb the tool without creating another layer of exception handling? These are not technical questions alone; they are operating questions. And in industrial businesses, the answers usually determine whether AI becomes margin expansion or just another IT expense.

In that sense, Welkin’s advantage is not that it “believes in AI” more strongly than others. It is that operating experience makes us less vulnerable to AI theater and more able to identify the businesses that can convert intelligence into better throughput, lower downtime, and smarter capital allocation. In a market where pilot purgatory remains a real risk,^[5] that distinction is not cosmetic. It is the difference between paying for optionality and underwriting compounding operating improvement.

Conclusion: the industrialist’s case for applied AI

The investment conclusion is straightforward: applied AI is becoming an operating system for repetitive industrial decisions, not a standalone product category. The firms most likely to capture durable value are not necessarily the ones building the best models, but the ones that own the workflow, control the process data, and have the management discipline to turn predictions into action. McKinsey’s recent work on operations leaders and manufacturing Lighthouses points to the same pattern: the organizations pulling ahead are the ones moving from pilots to scale, embedding AI across production networks, and treating deployment as a production capability rather than a one-off experiment.^[1][5][8]

For investors, that means the core question is not whether a company has “AI exposure.” It is whether AI can actually improve the economics of the business: lower downtime, reduce scrap and rework, raise throughput, speed customer response, improve forecast accuracy, and tighten capital allocation. Those outcomes depend on repeatable workflows, measurable bottlenecks, and leaders who can force the technology into decision loops with accountability.

The practical bias should therefore favor industrial businesses with three traits. First, they have dense operational data generated by their own assets, customers, or transactions. Second, they run standardized processes at enough scale that a marginal improvement compounds. Third, they can implement changes across maintenance, quality, procurement, service, and planning without getting trapped in pilot theater or organizational resistance.

Welkin’s operating heritage matters because it changes how those traits are underwritten. It improves pattern recognition in real workflows, sharpens skepticism toward exaggerated ROI claims, and makes it easier to distinguish between a compelling demo and a persistent earnings effect. In short: look for businesses that can convert intelligence into operating leverage. In this phase of AI, that is where the industrial economics live.